Example of a missed building Example of a false detection of building
Example of a poor performance of a detected building Figure 5. Final point cloud of buildings superimposed on an
orthoimage, using the bilateral filtering approach coloured in blue top and the scan line smooth
filtering approach coloured in yellow middle at the LIDAR point cloud; the CIR point cloud from SGM
coloured in magenta bottom.
For comparison reasons, a classification process utilizing the LIDAR point cloud is implemented using the software packages
of LAStools lasground_new, lasheight and lasclassify and ERDAS IMAGINE whose results are depicted in Figure 6.
Several combinations of parameters were tested for the classification process of the LAStools including search area
with a size of 1 m, 2 m, 3 m, 4 m and 5 m with the following combinations of building planarity and forest ruggedness:
0.1 m 0.2 m, 0.1 m 0.3 m, 0.1 m 0.4 m, 0.1 m 0.5 m, 0.2 m 0.4 m and 0.3 m 0.6 m. The best result was observed using a
search area of 5 m and a combination of building planarity and forest ruggedness of 0.3 m and 0.6 m respectively.
The parameters of the ERDAS IMAGINE that finally selected associated with the best results were:
Min slope = 30 deg, Min area = 5 m
2
, Min height = 2.5 m, Max area = 1400 m
2
, Plane offset = 0.3 m, Roughness = 0.6 m, Max height for low vegetation = 2 m,
Min height for high vegetation = 5 m. Also, several combinations of parameters were tested, e.g., the
Plane offset and Roughness took values: 0.2 m 0.5 m or 0.1 m 0.4 m or 0.3 m 0.5 m.
Example of a false detection of building Figure 6. Final point cloud of buildings superimposed on an
orthoimage, using the classification process in LAStools coloured in green top; the classification
process in ERDAS IMAGINE coloured in cyan bottom using the LIDAR point cloud.
3. EVALUATION OF THE RESULTS
The application area is a small town near Thessaloniki, in northern Greece, at an area of 0.33 Km
2
containing 501 industrial and residential buildings. The type of vegetation of
the scene is characterized as moderate. However, long arrays or groups of dense trees between the buildings, high vegetation
beside the boundary of buildings as well as buildings surrounded or occluded by high trees exist. This situation in
combination with the complex building structure with sloping roofs, chimneys, solar water heaters, small extensions or
additions of major buildings, etc, constitutes a challenge towards an accurate and reliable automatic building detection
process. To quantitatively evaluate the proposed approaches, the success rates of completeness, correctness and quality are
used. According to the ISPRS guidelines Rutzinger et al., 2009:
This contribution has been peer-reviewed. The double-blind peer-review was conducted on the basis of the full paper. Editors: S. Oude Elberink, A. Velizhev, R. Lindenbergh, S. Kaasalainen, and F. Pirotti
doi:10.5194isprsannals-II-3-W5-33-2015 37
TP Completeness=
TP FN
4
TP Correctness=
TP FP
5
TP Quality=
TP FP
FN
6 where TP, FP, and FN denote true positives, false positives, and
false negatives, respectively. Table 1 depicts the evaluation of the building results.
Approach Completeness
Correctness Quality
Scan line smooth
filtering 97.6
98.2 95.9
Bilateral filtering
99.8 96.0
95.8 CIR point
cloud 93.0
91.6 85.7
LAStools 92.0
79.2 74.1
ERDAS IMAGINE
99.0 84.1
83.4 Table 1. Evaluation of the automatic detection of building
points results. The scan line smooth filtering and the bilateral filtering
approaches achieved similar quality success rate. However, the scan line smooth filtering approach presented greater
correctness but less completeness than bilateral filtering approach due to its powerful filtering. Thus, although the
vegetation was almost completely removed, local complex cases of small extensions or additions of buildings which were
described partly due to the available density of the LIDAR point cloud were incorrectly removed increasing the FN. Conversely,
the bilateral filtering approach implements a more gently filtering and for this reason presents reverse performance on the
rates of correctness and completeness presenting more FP associated with cases of dense and high trees. Although the scan
line smooth filtering approach requires only the definition of the number of the symmetric points of the neighborhood, optimal
results using the same value of k for the filtering of the N
z
and roughness values were achieved. This is a comparative
advantage to the bilateral filtering approach as the parameters of spatial sigma and scalar sigma were differently tuned to achieve
the optimal results. On the other hand, point clouds that had been extracted by
dense image matching techniques suffer from other problems such occlusions, complex scenes, radiometric differences,
texture-less areas, etc. Thus, although the higher density of the CIR point cloud and the use of the NDVI which removed the
vegetation completely, FN and FP were observed mainly due to mismatches at complex cases of small buildings and unstable
interpolations respectively. Concerning the commercial software, the success rate of quality
of the LAStools is low mainly due to several false detections associated with the vegetation. Unlike the LAStools, the
ERDAS IMAGINE enables takes into account more parameters for the point classification process and therefore yielded higher
success rates.
4. EXPLOITATION OF THE PROPOSED
